Towards Fully Automated Systematic Reviews
In this post, I’m sharing my journey to build the ultimate paper summarizer agent — a lightweight, command-line tool powered by your favourite LLMs.
Writing a review paper sounded simple at first. Read a bunch of papers, extract key findings, summarize, and synthesize.
Right?
Wrong.
What I thought would be a straightforward academic task quickly became a time-consuming slog. Manually finding the right papers, understanding their content, and organizing them into a cohesive review was far harder than expected. It felt like I was reinventing the wheel every time I sat down to write.
That’s when I had an idea:
What if I built a tool that could do this grunt work for me?
- 🔍 Find relevant papers based on search query
- 🧹 Filter them using inclusion criteria
- 📖 Read and summarize each one into a structured format
- 🧾 Eventually help generate the scaffolding of a review paper
The Vision: A Researcher’s Dream Agent
I imagine the ultimate summarizer agent would work like this:
🔹 Input:
- A general research topic (e.g. “gene regulatory networks in synthetic biology”)
- Optional filters like date, species, or data type
- A preferred output format (CSV, table, markdown, etc.)
🔹 Output:
- A summary table of key metadata from each paper
- Optional Highlighted findings or quotes
- Possibly even draft review sections, organized by themes
- Talking with the body of papers
Phase 1: Systematic review Agent Using LangChain + OpenAI
To start, I tackled the hardest piece, and the most useful-one for me at least: screening papers and extracting relevant data from full-length PDF papers.
🛠 Tech Stack:
LangChain: to orchestrate chains and agentsPydantic: to validate at the field and model levelOpenAI GPT-4: to extract structured insights
🧑💻 Try It Yourself
I’ve open-sourced the full project here:
Key Features
- 🔍 Retrieve papers directly from PubMed Central
- 🧹 Screen abstracts with GPT-4o against your custom objectives
- 📊 Extract structured metadata (species, algorithms, results, etc.)
- 🧾 Export clean JSON/CSV summaries for analysis
🚀 Example Run
```bash research_agent analyze
–task meta
–objective “AI for diagnosing UTIs in dogs”
–filter-papers True
–max-results 50
–output-dir ./output